MORAIDICTIONNAIRE IA
Risques & Éthique

Differential Privacy

A mathematical guarantee that protects each individual by drowning their contribution in calibrated noise.

Differential privacy is a mathematical guarantee of privacy protection. Picture a crowd: adding or removing a single person leaves the group's overall silhouette unchanged. That is precisely the promise of this technique — the output of a data analysis should stay nearly identical whether or not any given individual is part of the dataset. No one can therefore infer whether your data was used.

How it works

The core idea is to inject calibrated random noise (often drawn from a Laplace or Gaussian distribution) into queries or models. This noise masks each person's contribution while preserving the overall statistical trends. We accept a slight loss of accuracy in exchange for a rigorous guarantee.

The guarantee is governed by a parameter ε (epsilon), the "privacy budget":

$$ \Pr[\mathcal{M}(D_1) \in S] \leq e^{\varepsilon} \cdot \Pr[\mathcal{M}(D_2) \in S] $$

where $D_1$ and $D_2$ are two datasets differing by a single individual. A low ε means strong protection (much noise); a high ε means weaker protection.

The central trade-off

Criterion Low ε High ε
Privacy Strong Weak
Noise added Heavy Light
Accuracy Reduced High

Differential privacy does not hide the tree — it makes the leaf that belongs to you impossible to recognize.

Explore the full AI dictionary →